Construction of an interpretable prediction model for poor functional outcome in conservatively managed basal ganglia hemorrhage based on CT radiomics and multiple machine learning algorithms.
Luo C, Xia J, et al. • Frontiers in neurology • 2026
An interpretable CT radiomics-based random forest model using 14 predictive variables—mostly wavelet radiomic features—can stably predict poor functional outcomes (mRS ≥ 4) at 3 months in patients with basal ganglia intracerebral hemorrhage receiving conservative treatment, achieving a bootstrap-corrected AUC of 0.867.
Key Findings
Results
A random forest model achieved the highest predictive performance among ten machine learning classifiers for predicting poor functional outcome in conservatively managed basal ganglia intracerebral hemorrhage.
Ten machine learning classifiers were trained and compared head-to-head, including a SuperLearner ensemble.
The random forest (RF) model delivered the highest AUC of 0.809, rising to a bootstrap-corrected AUC of 0.867 (95% CI, 0.797–0.929).
Model performance was further assessed via calibration curves, decision curve analysis, and the KS test.
The study enrolled 254 conservatively treated BG-ICH patients, randomly split into training and internal validation sets at a 7:3 ratio.
Results
After dimensionality reduction using univariate filtering and LASSO regression, 14 predictive variables were retained, predominantly consisting of wavelet radiomic features.
Radiomic features were extracted from admission plain (non-contrast) CT scans.
Univariate filtering followed by LASSO regression was used to isolate robust predictors.
The 14 retained variables were described as 'mostly wavelet radiomic features.'
Routine non-contrast CT intralesional density heterogeneity, previously overlooked by conventional clinical indicators, was captured through these radiomic features.
Results
Admission GCS score and key wavelet radiomic markers dominated prognostic importance in the random forest model.
SHAP (SHapley Additive exPlanations) analyses were used to unpack feature contributions and nonlinear links to continuous mRS scores.
Admission GCS score was identified as one of the top predictors alongside wavelet radiomic markers.
SHAP analyses provided interpretability for the model's predictions.
The outcome variable was defined as a 3-month modified Rankin Scale (mRS) score ≥ 4, signaling poor functional outcome.
Conclusions
The study was limited to internal validation and requires external multicenter cohort validation before clinical implementation.
The authors explicitly stated: 'At present, the model cannot be directly implemented in clinical practice, and external validation using independent multicenter cohorts is still required.'
The study was retrospective in design.
The training and validation split was conducted randomly at a 7:3 ratio within a single cohort of 254 patients.
Following adequate external cohort validation, the model 'may assist in individualized risk stratification and serve as a reference for subsequent clinical evaluations.'
Background
Conventional clinical indicators were found to largely overlook intralesional density heterogeneity visible on non-contrast CT, weakening long-term prognostic accuracy.
Routine non-contrast CT reveals intralesional density heterogeneity that is not captured by conventional clinical indicators.
This limitation of conventional approaches motivated the use of CT radiomics to extract quantitative features from admission scans.
Clinical variables were compared between poor and non-poor outcome groups as part of the analysis.
The incorporation of radiomic features alongside clinical variables aimed to improve prognostic accuracy beyond conventional methods.
What This Means
This research suggests that brain scans taken when a patient arrives at the hospital with a type of brain bleed called basal ganglia intracerebral hemorrhage contain hidden information that standard clinical assessments miss. By using a technique called 'radiomics,' which extracts hundreds of detailed mathematical features from CT scans—particularly 'wavelet' features that capture texture and density patterns invisible to the naked eye—the researchers were able to build a computer model that predicts which patients are likely to have poor physical function three months later. The model was built using data from 254 patients who were treated without surgery, and it correctly identified poor outcomes with an accuracy (AUC) of about 0.867, which is meaningfully better than conventional approaches.
The model was built using a 'random forest' algorithm, one of ten different machine learning approaches tested, and was made interpretable using a method called SHAP analysis, which shows how much each variable contributes to each prediction. The most important predictors were the patient's level of consciousness at admission (GCS score) and several wavelet-based radiomic features from the CT scan. This combination of clinical and imaging data outperformed models relying on clinical data alone.
This research suggests that quantitative CT imaging analysis could eventually help doctors better identify which hemorrhagic stroke patients managed without surgery are at highest risk of long-term disability, potentially allowing for more targeted monitoring and care planning. However, the authors emphasize that the model was only tested on one internal dataset and cannot yet be used in clinical practice—it must first be validated in independent patient groups from multiple hospitals before its real-world utility can be confirmed.
Luo C, Xia J, Qing R, Zeng Z, Deng X. (2026). Construction of an interpretable prediction model for poor functional outcome in conservatively managed basal ganglia hemorrhage based on CT radiomics and multiple machine learning algorithms.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1917496